Measuring genomic pre-selection in theory and in...
Transcript of Measuring genomic pre-selection in theory and in...
2007
Paul VanRaden and Jan Wright Animal Improvement Programs Lab, Beltsville, MD [email protected]
2013
Measuring genomic pre-selection in theory and in practice
Interbull annual meeting, Nantes, France, August 2013 (2) Paul VanRaden 2013
Traditional mixed models do not account for genomic selection • Phenotypes only for animals with
highest Mendelian sampling • GBV differ from EBV for progeny,
mates, parents, or herdmates
Multi-step methods may be biased Single-step methods reduce bias
In Theory
Interbull annual meeting, Nantes, France, August 2013 (3) Paul VanRaden 2013
1-Step Relationship Inverse Aguilar et al. (2010)
H-1 = A11 A12 A21 A22 + G-1 - A22
-1
1 = non-genotyped animals (60 million) 2 = genotyped animals (400,000)
Interbull annual meeting, Nantes, France, August 2013 (4) Paul VanRaden 2013
Animal model • EBV = w1 PA + w2 YD + w3 PC • (parent average, yield deviation,
progeny contribution)
1-step genomic info (GI) model • GBV = w1PAg + w2YDg + w3PCg + w4GI • GI = ∑off-diagonalj of G-1–A22
-1(GBVj) divided by diagonali of G-1–A22
-1 • Numerator of w4 in denominator of w
Traditional and 1-Step Models
Interbull annual meeting, Nantes, France, August 2013 (5) Paul VanRaden 2013
Computed for 8,300 Brown Swiss Diagonals of G and A22 are similar
• G: mean FG = 3.98% and SD = 4.15% • A22: mean FA = 3.95% and SD = 2.97%
Diagonals of G-1 larger than A22-1
• Mean = 5.83 for G, 2.18 for A22 • G-1, A22
-1 and difference all highly correlated
Diagonals of G-1–A22-1
Interbull annual meeting, Nantes, France, August 2013 (6) Paul VanRaden 2013
G A G-1 A-1 G-1 - A-1 G 1.0 .70 .05 .03 .06 A 1.0 .02 -.02 .04 G-1 1.0 .98 .99 A-1 1.0 .94 G-1 - A-1 1.0
Correlations of Diagonals Genotyped Brown Swiss
Interbull annual meeting, Nantes, France, August 2013 (7) Paul VanRaden 2013
Phenotyped only animals with good Mendelian sampling genotypes
Discrete or overlapping generations • Bias if discrete (Patry, Ducrocq 2011) • OK with overlap (Nielsen et al, 2012) • Large bias for dams (Liu et al, 2009)
Actual studies of pre-selection and genomic assortative mating needed
Simulations of Pre-Selection Bias
Interbull annual meeting, Nantes, France, August 2013 (8) Paul VanRaden 2013
Test actual selection and mating Quantify genomic pre-selection in:
• Mates of proven bulls (group 1) • Mates of young bulls (group 2) • Dams of young selected sons
Measure future bias because pre-selection has already occurred
In Practice
Interbull annual meeting, Nantes, France, August 2013 (9) Paul VanRaden 2013
Percentage of Genotyped Mates Group 2 bulls ranked by NM$
0
10
20
30
40
50
60
70
80
90
100
700 725 750 775 800 825 850 875 900 925
Maurice
Elvis ISY Altatrust
Fernand
Net Merit (Aug 2013)
Perc
enta
ge o
f mat
es g
enot
yped
Supersire
Numero Uno
S S I Robust Topaz
Garrold
Mogul
Interbull annual meeting, Nantes, France, August 2013 (10) Paul VanRaden 2013
Group 1 (proven bulls) • Daughters with records • Top 50, no dtrs in April 2010 • Were mates pre-selected?
Group 2 (young bulls) • Top 50, born 2009 and 2010 • Study calves born in USA • Will pre-selected mates cause bias?
Mates of Group 1 and 2 Bulls
Interbull annual meeting, Nantes, France, August 2013 (11) Paul VanRaden 2013
Trait Mean SD Min Max NM$ 2 4 -5 13 Protein 0 0 -1 1 Prod Life .0 .0 .0 .1 Dtr Preg Rate .0 .0 .0 .1 SCS .00 .00 -.01 .01 Final Score .00 .01 -.04 .02 Udder Depth .00 .01 -.02 .02
Group 1 Realized Mate Bias
Interbull annual meeting, Nantes, France, August 2013 (12) Paul VanRaden 2013
Trait Mean SD Min Max NM$ 8 9 0 33 Protein 0 0 0 1 Prod Life .1 .1 .0 .5 Dtr Preg Rate .1 .1 .0 .3 SCS -.01 .01 -.03 .00 Final Score .02 .03 -.01 .10 Udder Depth .03 .04 -.01 .13
Group 2 Future Bias from Mates
Interbull annual meeting, Nantes, France, August 2013 (13) Paul VanRaden 2013
Dams with ≥ 1 sampled son born 2008 to 2012
Son selection differential = ∑(GPTA – PA) / # of sons sampled
Dam’s bias = 2 * sons’ selection differential * (DE from sampled sons) / (total conventional DE) • DE = daughter equivalents or EDC
Future Bias – Dams of Young Bulls
Interbull annual meeting, Nantes, France, August 2013 (14) Paul VanRaden 2013
29 sons genotyped, 6 selected, each will provide 5.4 DE
Son selection differential for milk = ∑(GPTA – PA) / 6 = 583 pounds
30 daughters, each provide 1.5 DE 8.3 DE from PA, 7.8 from records Dam’s future bias = 2 * 583 * 6 * 5.4
/ [8.3 + 7.8 + 6 * 5.4 + 30 * 1.5] = 808
Example Dam HOUSA000065597532
Interbull annual meeting, Nantes, France, August 2013 (15) Paul VanRaden 2013
Trait Mean SD Min Max NM$ 29 33 -124 156 Protein 1 3 -10 14 PL .3 .5 -1.7 2.0 DPR .1 .2 -.9 .9 SCS -.01 .04 -.22 .14
Expected Future Bias – Bull Dams
Interbull annual meeting, Nantes, France, August 2013 (16) Paul VanRaden 2013
Also from preferential treatment • High-priced early daughters • Lack of random sampling
Deregression removes some bias • Example: dam gets credit only for
own records and non-genotyped progeny, not genotyped sons
• Use matrix instead of simple one at a time deregression
Potential Biases
Interbull annual meeting, Nantes, France, August 2013 (17) Paul VanRaden 2013
Conclusions
Evaluations should adjust for GBV instead of EBV of: • Progeny, mates, contemporaries, and
parents Biases from pre-selection:
• Very small for recently proven bulls • Moderate from mates top young bulls • Will be large for dams of several highly
selected sons, but deregression can remove some of the bias
Interbull annual meeting, Nantes, France, August 2013 (18) Paul VanRaden 2013
Phenotypes, genotypes, and pedigrees were provided by the Council on Dairy Cattle Breeding
Acknowledgments